Deep Adversarial Metric Learning.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Effective distance measurement is crucial for visual analysis, often relying on hard negative samples.
• Hard negatives are typically a minority, potentially limiting the description of data distributions near decision boundaries.

Purpose of the Study:

• To develop a deep adversarial metric learning (DAML) framework for generating synthetic hard negatives.
• To enhance existing supervised deep metric learning algorithms by exploiting easy negatives.

Main Methods:

• DAML framework simultaneously trains feature embedding and a hard negative generator adversarially.
• Deep adversarial multi-metric learning (DAMML) learns multiple local transformations for comprehensive input space description.
• Exploits collaborative and competitive relationships among multiple metrics, using a metric discriminator to prevent overlap.

Main Results:

• DAML and DAMML frameworks generate targeted synthetic hard negatives.
• Adversarial training leads to more precise distance metrics.
• Experimental results demonstrate performance boosts on benchmark datasets.

Conclusions:

• DAML and DAMML effectively enhance existing deep metric learning approaches.
• Adversarial learning provides a powerful mechanism for improving distance metric accuracy.
• The proposed methods offer a widely applicable solution for visual analysis tasks.